开源AI工作流

Install

Supported platforms: macOS and Linux. Windows is not supported because hm uses Unix process exec semantics and Unix filesystem permissions for launch, isolation, and secret handling.

curl -fsSL https://raw.githubusercontent.com/INONONO66/harness-manager/main/scripts/install.sh | sh

Install and copy built-in manifests in one step:

curl -fsSL https://raw.githubusercontent.com/INONONO66/harness-manager/main/scripts/install.sh | sh -s -- --init

Install everything non-manual that hm init --install can manage:

curl -fsSL https://raw.githubusercontent.com/INONONO66/harness-manager/main/scripts/install.sh | sh -s -- --install-harnesses

Or build from source:

git clone https://github.com/INONONO66/harness-manager.git
cd harness-manager
cargo build --release
cp target/release/hm ~/.local/bin/

For local development, install from the checkout:

cargo install --path .

Preview injected endpoint, bearer, and proxy env

hm inject plan codex --profile proxy

Launch a runtime with a clean profile env

hm use codex --profile proxy hm use claude --profile proxy

Injection strategy 1: `env` (single-provider runtimes)

[injection]
strategy = "env"
provider = "anthropic"
supported_providers = ["anthropic"]
endpoint_env = "ANTHROPIC_BASE_URL"
api_key_env = "ANTHROPIC_API_KEY"
strip_envs = ["ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_BASE_URL"]
endpoint_strip_v1 = true

hm use claude --profile proxy strips strip_envs from the child env, then sets endpoint_env and api_key_env from the active profile's gateway. If no gateway is present and the legacy [profiles.X.llm] block is, hm falls back to that single endpoint/bearer.

Injection strategy 2: `provider-config-seed` (multi-provider runtimes)

[injection]
strategy = "provider-config-seed"
config_path = "{home}/.config/opencode/opencode.json"
root_key = "provider"
provider_base_url_key = "options.baseURL"
provider_api_key_key = "options.apiKey"
provider_headers_key = "options.headers"
supported_providers = ["anthropic", "openai", "google", "openrouter", "groq", "xai", ...]
overwrite = false
endpoint_strip_v1 = false
legacy_provider = "openai"

[injection.provider_api_key_envs]
anthropic = "ANTHROPIC_API_KEY"
openai = "OPENAI_API_KEY"
google = "GOOGLE_API_KEY"

[injection.provider_header_overrides.anthropic]
"x-api-key" = "{bearer}"
"Authorization" = "Bearer {bearer}"

hm use opencode --profile proxy writes a JSON file under the isolation home (never the user's real ~). The file is deep-merged into any existing user content. legacy_provider (optional) tells the strategy how to fall back to [profiles.X.llm]: hm seeds that one provider with llm.endpoint and llm.bearer. provider_api_key_envs maps each supported provider id to the child-process env var that should receive the resolved bearer; missing mappings fail closed before env or file writes.

Injection strategy 3: `codex-config-seed` (top-level TOML + env hybrid)

[injection]
strategy = "codex-config-seed"
config_path = "{home}/.codex/config.toml"
openai_base_url_key = "openai_base_url"
model_provider_key = "model_provider"
model_provider_value = "openai"
provider = "openai"
supported_providers = ["openai"]
api_key_env = "CODEX_API_KEY"
strip_envs = ["OPENAI_API_KEY", "OPENAI_BASE_URL", "CODEX_API_KEY", "CODEX_ACCESS_TOKEN"]
overwrite = false
endpoint_strip_v1 = false

hm use codex --profile proxy writes two top-level keys to the TOML config (via toml_edit::DocumentMut, preserving comments and existing keys from [isolation.seed_files]), then strips strip_envs from the launch env and sets api_key_env to the resolved bearer. The bearer NEVER reaches the file — only the env var. This strategy is single-provider only: the gateway must route the configured provider (or the legacy [profiles.X.llm] fallback is used).

Adding a new runtime requires only a TOML file. No Rust change.

bundled and user/plugin harness manifests are loaded through the same registry

```